CN116449305A

CN116449305A - Dense time-varying array construction method and system based on controllable variation self-encoder

Info

Publication number: CN116449305A
Application number: CN202310426246.4A
Authority: CN
Inventors: 原达; 孙文力
Original assignee: Shandong Technology and Business University
Current assignee: Shandong Technology and Business University
Priority date: 2023-04-17
Filing date: 2023-04-17
Publication date: 2023-07-18

Abstract

The invention discloses a dense time-varying array construction method and system based on a controllable variable self-encoder, and belongs to the technical field of ground penetrating radar data processing. The invention provides a constraint embedded dense time-varying array construction method based on controllable variable self-coding, which uses a time-varying data reconstruction network based on a controllable variable self-coder to generate pseudo-channel data close to real distribution to increase array density through hidden space sampling interpolation and fusion of depth layer characteristic information; the data registration module based on scale invariant feature transformation is constructed, the inter-track gradient features and the structural symmetry features are extracted, and the data registration can be completed in a space-time domain; and meanwhile, the shallow layer features are used as constraint information to be embedded into a data reconstruction network, so that the influence of data channel offset on array generation is eliminated. The method solves the problem that in the prior art, in the time-varying data interpolation reconstruction process, reconstruction errors are accumulated continuously, so that the quality of generated data is reduced.

Description

Dense time-varying array construction method and system based on controllable variation self-encoder

Technical Field

The invention relates to the technical field of ground penetrating radar data processing, in particular to a dense time-varying array construction method and system based on a controllable variable self-encoder.

Background

The statements in this section merely relate to the background of the present disclosure and may not necessarily constitute prior art.

Ground penetrating radar (Ground Penetrating Radar, GPR) is an electromagnetic technique for detecting the distribution of an underground medium by electromagnetic waves and determining the position of the internal structure or shape thereof, and has been widely used in many fields such as underground structure surveying in recent years. In a wide range of geological survey applications, the use of GPR data to construct high density arrays, forming three-dimensional visual models with continuous physical structures, can provide multiple viewing perspectives, enhancing the reliability of the analysis.

The data collected by the ground penetrating radar is reflection data obtained according to electromagnetic scattering fields, and three forms of A-Scan, B-Scan and C-Scan are commonly adopted. In existing three-dimensional subsurface structure analysis applications, building a high-density time-varying array based on two-dimensional profile data (B-scan) is one of the main approaches. However, when constructing three-dimensional time-varying arrays using B-scan, the following problems have not been solved:

(1) The construction of three-dimensional dense time-varying arrays requires first high quality time-varying data, and the main strategy in the currently existing methods is to gradually add new track data (a-scan) from registration and reconstruction of successive B-scan arrays until the entire sequence is reconstructed. However, in performing a wide range of geological surveys, the acquired B-scan array is subject to variations in the parameters of the terrain, survey lines and survey conditions, resulting in sparse spatial data. The slice interval of the B-scan data set is irregular, the inter-slice data density is far lower than the inter-track data density, trace loss is easy to generate in the process of constructing the three-dimensional model, and the spatial structure of the three-dimensional model is deformed. Therefore, data generation is required for the space area not covered by the scanning wires, and the data density between slices is improved.

(2) The acquisition of the B-scan data set is not synchronous, irregular offset exists among time-varying data due to the difference of time zero positions, and in the time-varying data interpolation reconstruction process, reconstruction errors are accumulated continuously, so that the quality of generated data is reduced. Therefore, data registration of the measured data is required to eliminate irregular offset between the data.

For the reconstruction method of the GPR dense time-varying array, in short, the spatial data density is improved, the space-time correlation among the channel data is utilized to interpolate the existing data sequence, and the sparse array is interpolated to achieve the purpose of constructing the dense array. Currently, the main strategy for time-varying data interpolation algorithms is to interpolate the effective data gap using a moving average of two adjacent effective values of sparse spatial data and to utilize the spatio-temporal correlation of the surrounding data of the sparse spatial data.

In the prior art, some students use cubic spline interpolation to reconstruct time-varying data, and any two information points are connected by a straight line or a smooth curve. However, GPR data is nonlinear time-varying data with strong fluctuations, and this method masks local feature variations of adjacent interpolation points, so that the effect of GPR time-varying data is not ideal. Some scholars reconstruct time-varying data based on time-series harmonic analysis (Harmonic Analysis of Time Series, HANTS), and use a least squares curve fitting method based on harmonic components, demonstrating good performance in time-series interpolation. Some students use singular spectrum analysis (singular spectral analysis, SSA) to iteratively estimate the data gap, optimizing the window width and the number of SSA primary modes through cross-validation to fill in the sparse data gap. The algorithm is obtained based on a mathematical model, and has good generalization performance in application. However, because of the lack of spatial reference information in its approach, the global consistency of the results of interpolating the extensive survey data is not high. Some scholars utilize the frequency spectrum and the space information simultaneously, and an analysis method for reconstructing lost data in a time-varying signal based on a sparse representation method is provided. Some students use the DAISY feature descriptors for data registration to generate dense spatial data, and then extract candidate plane sets on the basis to construct a dense time-varying array. Such methods can reference more information in multiple dimensions, but the time cost in interpolating large area missing data based on spatial information is generally high. Furthermore, trace data can vary widely over short distances, resulting in such methods not being able to characterize real data.

In recent years, students at home and abroad integrate the emerging deep learning technology into the existing model while keeping the research on the prior method, and innovatively optimize the existing model. Recently, sparse data interpolation deep learning neural networks of different structures such as GRU, LSTM, GAN, VAE are emerging. In the application of data interpolation, the two structural neural networks, namely GRU (gate recurrent unit) and LSTM (long short-term memory), can reserve important features through different thresholds, ensure that the important features are not lost in the long-term transmission process, learn rules from observed data, learn further from the interpolated data, and improve the data interpolation precision. However, the GRU and LSTM have limitations, and the GRU has advantages only in the scenes of long time-varying data and small sample sets; LSTM still has a gradient that disappears when faced with sequences beyond the length limit, and requires that the sequence data be strictly time-dependent and not applicable to GPR data where there is a time zero problem. GAN (generative adversarial networks) networks are used to learn the overall distribution of time-varying data, thereby interpolating the gaps of sparse data. However, because of the fixed information content, when the GAN network generates complex data such as GPR time-varying data, the generated data has poor spectrum or time-spectrum performance, and the interpolation precision is not high, the model is unstable, and the like. Kingma equals 2014, proposes a variable self-encoder (Variational Autoencoder, VAE). The VAE comprises an encoder and a decoder, wherein the encoder deduces input data into a variation probability distribution conforming to a normal distribution; the decoder restores the variational probability distribution to approximate the approximate probability distribution of the original data, and forcibly learns the continuous hidden space distribution by adding KL divergence to the loss function. However, VAEs increase in divergence during data generation, thereby resulting in reduced feature learning ability and blurring of output. The Denton et al proposed-VAE to address this problem by increasing the divergence weight, artificially reducing the divergence to give it better learning ability, at the cost of higher reconstruction errors, than the VAE, although the divergence is adjusted by the control system. Based on-VAE, shao et al propose controlling VAE, by introducing a nonlinear proportional integral derivative controller to adapt to super parameters, the mutual information between the reconstructed data and the hidden space is maximized, and the balance of reconstruction error and divergence weight distribution is realized. However, as an unsupervised learning method, when a time-varying sequence is generated using a control vae, a matching relationship of an input sequence cannot be captured, and accurate interpolation of sparse data is realized.

Disclosure of Invention

In order to solve the defects in the prior art, the invention provides a dense time-varying array construction method, a system, electronic equipment and a computer readable storage medium based on a controllable variable self-encoder, which can better distribute weights of balance reconstruction errors and KL divergence, improve the capability of learning features of a network model and the quality of reconstructing hidden space data, eliminate irregular offset by using constraint information and finally generate the dense time-varying array.

In a first aspect, the present invention provides a dense time-varying array construction method based on a controllable variable self-encoder;

the method for constructing the dense time-varying array based on the controllable variable self-encoder comprises the following steps:

acquiring two-dimensional profile data, wherein the two-dimensional profile data is acquired by a ground penetrating radar;

based on the region joint feature constraint, performing time-varying data registration on the track data in the two-dimensional profile data to acquire additional feature information;

inputting the additional characteristic information and the track data into a preset controllable variable self-encoder for processing, and obtaining a reconstructed dense time-varying array;

the controllable variation is used for extracting deep characteristic information of the track data from the encoder, and carrying out cyclic iteration by taking the additional characteristic information as shallow characteristic constraint to obtain a dense time-varying array.

Further, based on the region joint feature constraint, performing time-varying data registration on the track data in the two-dimensional profile data, and acquiring additional feature information includes:

partitioning the two-dimensional section data according to the displacement direction to obtain one-dimensional channel data;

calculating the gradient direction and gradient amplitude of adjacent one-dimensional channel data, and sequentially aligning corresponding channel data in continuous one-dimensional channel data according to the gradient direction and the gradient amplitude;

using a 9 multiplied by 9 square neighborhood as a data gradient statistical range, and obtaining a feature vector by respectively calculating gradient direction and gradient amplitude accumulated values of each point so as to perform feature matching to obtain a two-dimensional profile data array;

and calculating correlation coefficients of the reference channel and the registration channel in the two-dimensional profile data array in sequence, regarding two one-dimensional channel data corresponding to the maximum value of the coefficients as the most approximate channel, and taking the position, the amplitude and the main direction of the most approximate channel as additional characteristic information.

Further, the inputting the additional feature information and the track data into the preset controllable variable self-encoder for processing comprises:

rolling and subsampling the input track data, and compressing the data into a low-dimensional vector; normalizing the input additional feature information to be used as a constraint condition vector;

Inputting the low-dimensional vector and the constraint condition vector into an encoder, carrying out joint distribution sampling on the low-dimensional vector and the constraint condition vector, generating a hidden vector containing shallow information and deep features, and projecting the hidden vector into a hidden variable space similar to the constraint;

interpolation sampling is carried out between hidden vectors in the two hidden spaces to form a time-varying sequence synthesized vector;

inputting the time-varying sequence synthesized vector into a decoder to generate reconstruction channel data; and carrying out hidden space interpolation reconstruction on all other hidden vectors to obtain a complete dense time-varying array.

Preferably, the performing joint distribution sampling on the low-dimensional vector and the constraint condition vector specifically includes: inputting the low-dimensional vector and the constraint condition vector into an encoder, performing multi-element Gaussian distribution processing on the low-dimensional vector and the constraint condition vector, decomposing the input vector into a mean vector and a standard deviation vector which accord with multi-element Gaussian distribution, multiplying the standard deviation by random noise which accord with standard normal distribution by using a re-parameterization skill, and adding the mean to obtain normal distribution of reconstruction parameters; sampling from the reconstruction parameters and splicing with the condition vector to obtain the hidden vector containing both shallow information and deep features.

Preferably, the interpolation sampling between the hidden vectors in the two hidden spaces forms a new time-varying sequence synthesized vector specifically includes: establishing a mapping between points of a priori distribution area in the hidden space and the input, wherein the mapping is realized through reconstruction loss of a decoder and hidden variable weight parameter constraint; and then interpolating and sampling between hidden vectors in the two hidden spaces to form a time-varying sequence synthesized vector.

Furthermore, in the training process of the controllable variable self-encoder, the KL divergence of each training step is sampled by using a feedback control idea, and the super parameter beta (t) is correspondingly adjusted through an iterative algorithm so as to enable the KL divergence to be stabilized at a desired value.

Further, the controllable variable self-encoder comprises a convolution layer, a pooling layer, an encoder and a decoder which are connected in sequence.

In a second aspect, the present invention provides a dense time-varying array building system based on a controllable variable self-encoder;

the first acquisition module is used for acquiring two-dimensional profile data, wherein the two-dimensional profile data is acquired by a ground penetrating radar;

the second acquisition module is used for carrying out time-varying data registration on the track data in the two-dimensional section data based on the region joint feature constraint to acquire additional feature information;

the dense time-varying array acquisition module is used for inputting the additional characteristic information and the track data into a preset controllable variable self-encoder for processing to acquire a reconstructed dense time-varying array;

the controllable variation is used for extracting deep characteristic information of the track data from the encoder, and carrying out cyclic iteration by taking the additional characteristic information as shallow characteristic constraint to obtain a dense time-varying array. In a third aspect, the present invention provides an electronic device;

An electronic device comprising a memory and a processor and computer instructions stored on the memory and running on the processor, which when executed by the processor, perform the steps of the above described method of dense time varying array construction based on a controllable variable self-encoder.

In a fourth aspect, the present invention provides a computer-readable storage medium;

a computer readable storage medium storing computer instructions which, when executed by a processor, perform the steps of the dense time varying array building method based on a controllable variable self-encoder described above.

Compared with the prior art, the invention has the beneficial effects that:

the invention provides a method for constructing a constrained embedded dense time-varying array based on a controllable variable self-coding network, which utilizes the characteristic that hidden space features of a controllable variable self-coding network are continuous to sample and interpolate and reconstruct hidden space data, interpolates data in a sparse region and improves the space data density. And dynamically adjusting the loss weight through a controller, and balancing the generation quality and the feature learning capacity of the decoder until the generation precision approaches to the real data. Data registration is then performed using a modified PSO-SIFT algorithm. And the time-varying data are subjected to joint registration in a space domain and a time domain, constraint information is extracted by designing feature descriptors which accord with the structural features and gradient changes of the GPR time-varying data, and the influence of irregular offset on data reconstruction is effectively eliminated.

Experimental results show that compared with a mainstream algorithm, the method provided by the invention has higher robustness, the generated dense time-varying array has natural transition, and the method can effectively represent the real underground environment in the face of real data with complex structure and multiple interferences.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention.

FIG. 1 is a schematic flow chart of a dense time-varying array construction method based on a controllable variable self-encoder according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a three-dimensional GPR time-varying data array field provided by an embodiment of the invention;

FIG. 3 is a schematic diagram of a feature descriptor in the form of a loop according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of B-scan spatial domain registration results provided by an embodiment of the present invention;

FIG. 5 is a schematic diagram of an A-scan time domain registration result according to an embodiment of the present invention;

FIG. 6 is a schematic diagram of a dynamic feedback calculation flow of a controller according to an embodiment of the present invention;

FIG. 7 is a schematic diagram of a processing procedure of input data and condition vectors according to an embodiment of the present invention;

FIG. 8 is a schematic diagram of a hidden space interpolation reconstruction flow provided in an embodiment of the present invention;

FIG. 9 is an enlarged schematic diagram of the interpolation results and the array of the areas a and b according to the embodiment of the present invention;

fig. 10 is a schematic diagram of a zoom-in and zoom-out comparison of a region c through GAN interpolation according to an embodiment of the present invention;

FIG. 11 is a schematic diagram of the generation result of a GPR dense time-varying array according to an embodiment of the present invention;

FIG. 12 is a schematic diagram showing the three-dimensional reconstruction result and the cross section of an original sparse time-varying array according to an embodiment of the present invention;

FIG. 13 is a schematic diagram showing the three-dimensional reconstruction result and cross-section of a dense time-varying array according to an embodiment of the present invention;

fig. 14 is a schematic diagram of a three-dimensional reconstruction result of a constraint embedding ablation experiment according to an embodiment of the present invention;

fig. 15 is a schematic diagram of a hyperbolic crest calibration visualization result provided by an embodiment of the present invention.

Detailed Description

It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the invention. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms also are intended to include the plural forms, and furthermore, it is to be understood that the terms "comprises" and "comprising" and any variations thereof are intended to cover non-exclusive inclusions, such as, for example, processes, methods, systems, products or devices that comprise a series of steps or units, are not necessarily limited to those steps or units that are expressly listed, but may include other steps or units that are not expressly listed or inherent to such processes, methods, products or devices.

Embodiments of the invention and features of the embodiments may be combined with each other without conflict.

Example 1

The embodiment provides a dense time-varying array construction method based on a controllable variable self-encoder, and aims to solve the problems of low spatial data density and irregular data offset in the dense time-varying array construction process. And the regional division and multi-domain combined thought is adopted to register the track data and generate a characteristic descriptor, and an embedded data reconstruction algorithm is restrained, so that the accumulated error caused by irregular data offset is effectively eliminated. The loss function of the control VAE is dynamically adjusted, the balance of reconstruction errors and KL divergence weight distribution is realized, and the quality of generated data is improved. Sampling and interpolating reconstruction in the hidden space, and fusing multi-level characteristic information to generate reconstruction data with continuous characteristics.

Next, a detailed description will be given of a dense time-varying array construction method based on a controllable variable self-encoder disclosed in this embodiment with reference to fig. 1 to 15. The method comprises the following specific steps:

s1, acquiring two-dimensional profile data, wherein the two-dimensional profile data are acquired by a ground penetrating radar. And carrying out time-varying data registration on the track data in the two-dimensional profile data based on the region joint feature constraint, and obtaining additional feature information. Comprises the following steps of;

(1) Partitioning the two-dimensional section data according to the displacement direction to obtain one-dimensional channel data;

(2) Calculating the gradient direction and gradient amplitude of adjacent one-dimensional channel data, and sequentially aligning corresponding channel data in continuous one-dimensional channel data according to the gradient direction and the gradient amplitude;

(3) Using a 9 multiplied by 9 square neighborhood as a data gradient statistical range, and obtaining a feature vector by respectively calculating gradient direction and gradient amplitude accumulated values of each point so as to perform feature matching to obtain a two-dimensional profile data array;

(4) And calculating correlation coefficients of the reference channel and the registration channel in the two-dimensional profile data array in sequence, regarding two one-dimensional channel data corresponding to the maximum value of the coefficients as the most approximate channel, and taking the position, the amplitude and the main direction of the most approximate channel as additional characteristic information.

Specifically, the imaging principle of GPR is to transmit a detected electromagnetic pulse into the subsurface and receive a corresponding echo signal having a time-varying characteristic. From the perspective of time-varying array analysis, a C-scan consisting of an arrangement of multiple parallel time-varying signals is considered a three-dimensional time-varying array, as shown in FIG. 2.

When the detection pulse encounters the target object, the echo changes obviously, and is reflected as a hyperbolic wave in the B-scan data. As the GPR data are acquired and influenced by factors such as time zero point, measuring environment and the like, the hyperbolic wave in the B-scan is irregularly offset, and the same target object shows different hyperbolic wave structures. Therefore, the track data in the B-scan is registered, and the accumulated error generated by reconstruction is reduced. The raw B-scan data can be expressed as:

B-scan＝X(i,j)

Wherein i=1, 2,3,..n; j=1, 2,3,..m.

The method comprises the steps of dividing the displacement direction into blocks, wherein the obtained one-dimensional A-scan data are as follows:

W _j ＝X _j [1,2,...,n]＝{W _j (1),W _j (2),W _j (3),...W _j (n)}

since GPR data has complex correlations in both the spatial and temporal domainsThe present embodiment therefore registers the a-scan data from two sides. In the spatial domain, since data acquisition generally requires designing a plurality of parallel lines, it is difficult to determine the starting position of each line. Thus requiring data stitching of consecutive B-scan data at horizontal spatial locations. Because the relative positions of the array antennas are not changed, the section initial positions of the B-scan data are always equal, and corresponding track data in continuous B-scan data can be aligned in sequence by calculating the gradient direction and gradient amplitude of one-dimensional track data in adjacent B-scan data as registration basis. Will W _j (i) Is defined as:

G _x ＝([W _j+1 (i-1)+2W _j+1 (i)+W _j+1 (i+1)]

-[W _j-1 (i-1)+2W _j-1 (i)+W _j-1 (i+1)])

G _y ＝([W _j-1 (i-1)+2W _j (i-1)+W _j+1 (i-1)]

-[W _j-1 (i+1)+2W _j (i+1)+W _j+1 (i+1)])

wherein G is _x Represented at X _j (i) Horizontal derivative of G _y Representing its vertical derivative. Classical SIFT algorithms use the correlation between features to determine the optimal registration relationship, but when SIFT is used directly for trace data registration, the correct number of correspondences is insufficient to confirm the matching accuracy due to the significant differences in echo signal intensity mapping.

Therefore, the present embodiment combines the improved PSO-SIFT algorithm region division concept, fully uses the continuous correlation and the spatial structure characteristic between the trace data in the B-scan array, designs the character-back feature descriptor as shown in fig. 3, and the improved feature descriptor contains more neighborhood information.

Because the target area in the B-scan data shows a hyperbolic wave structure, the method has higher structural symmetry, and compared with the original SIFT algorithm, the area division concept has better advantages on the irregular offset phenomenon of the track data, and the added neighborhood information enables the descriptor to be more stable. Using a 9×9 square neighborhood as a data gradient statistical range, feature vectors with dimensions of 9×8=72 dimensions are sequentially obtained by calculating gradient directions and gradient magnitude accumulated values of each point, respectively. Finally, registration is completed through multiple feature matching, the registration effect is shown in fig. 4, and the corresponding relation of the track data in the B-scan array is clearly embodied.

In the time domain, since the transmission and reception distances of the array antennas are equal and cannot change, the starting points of the continuous A-scan data are ideally identical. Calculating correlation coefficient C of a reference track and a registration track in the B-scan array in sequence, and setting reference track data as W _r Registration trace data is W _c C can be expressed as:

taking two paths of A-scan data corresponding to the maximum value of the coefficient as the most approximate path, sampling the data as the input of a subsequent data generation model, and generating the pseudo-path data through interpolation calculation. Respectively extracting two registration track sets W= (W) in a continuous B-scan array ₁ ,W ₂ ,...W _N ) And W' = (W ₁ ′,W ₂ ′,...W _N '). With (x, y), σ, θ and (x ', y '), σ ', θ ' denote the position, magnitude and principal direction of the registration tracks W and W ', respectively, in the corresponding B-scan data. The matching result is shown in fig. 5.

In the subsequent step, three types of information, namely (x, y), sigma and theta, between the extracted feature point pairs are reserved as additional feature information F in a data reconstruction model, and are spliced with deep feature information extracted from the controllable VAE as hidden vectors Z as shallow feature constraints in a cyclic iteration process. The spliced hidden vector not only has hidden space characteristics of higher level abstraction, but also has shallow layer characteristics such as space position, amplitude, main direction and the like, and can effectively eliminate irregular offset among data when a decoder reconstructs the data, thereby improving the accuracy of reconstructing an array field.

S2, inputting the additional characteristic information and the track data into a preset controllable variable self-encoder for processing, and obtaining a reconstructed dense time-varying array. The method comprises the following steps:

(1) Rolling and subsampling the input track data, and compressing the data into a low-dimensional vector; normalizing the input additional feature information to be used as a constraint condition vector;

(2) Inputting the low-dimensional vector and the constraint condition vector into an encoder, carrying out joint distribution sampling on the low-dimensional vector and the constraint condition vector, generating a hidden vector containing shallow information and deep features, and projecting the hidden vector into a hidden variable space similar to the constraint;

(3) Interpolation sampling is carried out between hidden vectors in the two hidden spaces to form a time-varying sequence synthesized vector;

(4) Inputting the time-varying sequence synthesized vector into a decoder to generate reconstruction channel data; and carrying out hidden space interpolation reconstruction on all other hidden vectors to obtain a complete dense time-varying array.

Next, the process of the controllable variable self-encoder will be described in detail.

(1) Controllable variable-division self-encoder

For the problem of VAE inability to balance reconstruction loss and KL divergence, the present embodiment introduces a controllable variable self-encoder (ControlVAE) that stabilizes the value of KL divergence by dynamically adjusting the weight of KL divergence. Inspiring by a control system, the control VAE samples the KL divergence of each training step by using a feedback control idea, and correspondingly adjusts the super parameter beta (t) through an iterative algorithm to enable the KL divergence to be stabilized at an expected value.

The controllable variable self-encoder is trained, the control VAE is trained twice, and the posterior distribution of the real data is approximated for the training decoder for the first time, so that high-quality reconstruction data is generated. And mapping the hidden vector of the hidden space interpolation sampling back to the channel data for the second time, thereby improving the space data density. Wherein the hidden vector is embedded with constraint information, thereby weakening the irregular offset phenomenon.

During model training, sampling KL divergence output in the training step t by using V _K ′ _L (t) represents. Finding the sampled KL divergence and the set desired V _KL Is a difference e (t) of the following formula:

e(t)＝V _KL -V _K ′ _L (t)

the super parameter β (t) is calculated using e (t) as feedback, where the lower variation bound function of the ControlVAE is:

L _conVAE ＝E _Q(z|x) [logP _θ (x∣z)]-β(t)KL(Q _Φ (z∣x)‖P _θ (z))

wherein the super parameter β (t) is defined as:

K _p and K _i Is a constant set by a priori knowledge. When the KL divergence falls below the desired value, the controller slows down the decline of the KL divergence by reducing the hyper-parameter β (t), i.e. reducing the penalty on the KL divergence in the varying lower bound function. At the same time, the decreasing hyper-parameter β (t) increases the varying lower bound function and thus increases the KL divergence until the desired value is exceeded. In the same way, when the KL divergence exceeds the desired value, the controller decreases it by increasing the hyper-parameter β (t), i.e. increasing the penalty on the KL divergence in the varying lower bound function.

The first term of the super parameter beta (t) dynamically influences the value of the super parameter beta (t) according to the value of the difference e (t), and promotes the KL divergence to be dynamically increased or decreased in a favorable direction. The second term of the super parameter beta (t) sums the accumulated differences generated by the training samples to produce a progressively stronger correction until the sign of the difference changes. This ensures that β (t) gradually changes toward a direction that favors the KL divergence approaching the desired value. The calculation flow of the super parameter beta (t) is shown in fig. 6.

(2) Constraint embedding hidden vector construction

For the generation of GPR pseudo-channel data, if the hidden vectors input to the decoder contain enough characteristic information, not only the structural information in the original data can be effectively reserved, but also the irregular offset phenomenon between the data can be weakened as much as possible. Therefore, the embodiment designs a constraint embedded control VAE to perform feature extraction and constraint reconstruction on the A-scan data, and constructs the hidden vector containing multiple feature information.

In the feature extraction work, the convolution layer and the pooling layer of the control vae perform convolution and subsampling on the input track data, compress the data into a low-dimensional vector S, perform normalization processing on the position (x, y), the amplitude sigma and the main direction θ corresponding to the registration track set extracted in the second section, and enable f= ((x, y), sigma and θ) to be used as constraint condition vectors for synchronous input. The compressed low-dimensional vector S and the conditional vector are input into an encoder, the encoder is subjected to multi-element Gaussian distribution processing, the input vector is decomposed into a mean vector mu and a standard deviation vector lambda which accord with multi-element Gaussian distribution, and a re-parameterization skill is used, and the standard deviation is multiplied by random noise e which accords with standard normal distribution _i And adding the average value to obtain the normal distribution h of the reconstruction parameters. Finally, sampling from the reconstruction parameters and splicing with the condition vector to obtain a hidden vector Z which simultaneously contains shallow information and deep features, as shown in fig. 7. The hidden vector accords with normal distribution and is generated by jointly distributing and sampling input data and the conditional vector.

(3) Continuous hidden space interpolation

In the training process, the encoder inputs two continuous A-scan at a time, generates two corresponding hidden vectors and projects the corresponding hidden vectors into a hidden variable space similar to the constraint. The decoder inputs one hidden vector at a time and generates one reconstructed track data. After training is completed, the decoder can map features in the hidden feature space back to the reconstructed a-scan data as much as possible. Then we use the control vae to sample, which is equivalent to scattering the basic units of the two data, and the encoder creates a mapping between the points of a prior distribution area in the hidden space and the input, this mapping is achieved by the decoder reconstructing the loss and under the constraint of hidden variable weight parameters. A new time-varying sequence synthesized vector Znew is formed by interpolation sampling between vectors in the hidden space corresponding to the two input data, and then the false channel data Snew in the space region is reconstructed by a decoder. The interpolation process is shown in fig. 8, where the input of Sq and Se to the encoder simultaneously results in two hidden vectors Zq and Ze. Since the latent feature space has continuity, interpolation in the latent feature space is reasonable. The track data Snew generated by decoder reconstruction and the corresponding hidden feature vector Znew are derived from the following equation:

Znew＝Zq+rand(0,1)(Ze-Zq)

Snew＝decoder(Znew)

And reconstructing the spatial interpolation between the registration channel sets selected by registration once to obtain hidden space artifact data. After the hidden space interpolation reconstruction is carried out on all other registration tracks, a complete dense time-varying array is obtained.

Next, the advancement of the dense time-varying array construction method based on the controllable variable self-encoder described in this embodiment was verified in connection with experiments.

(1) Experimental data set

In the embodiment, 3 groups of GPR simulation data and 7 groups of real data obtained by detecting underground buried pipelines are used for verifying and analyzing the method, wherein the scanning gap of the 7 groups of real data is 1 meter; 3 groups of simulation data are obtained by GPRMAX simulation, and the data gap is 0.5 m. These B-scan slices make up 10 sets of sample data for training a control vae based time-varying data generation network. In order to reduce redundant characteristics and operation time during network training, the original data is uniformly cut into 512 pixels by 512 pixels.

(2) Training details and evaluation index

The hardware platform of the experiment is Intel (R) Core (TM) i7-10700 CPU@2.90GHz memory 32GB RAM, and an image processor (GPU) adopts NVIDIA GTX2070. During training, K is taken _p And K _i And respectively setting the values to be 0.01 and 0.001, sequentially inputting a pair of sequence B-scan data for training, adjusting the structure and parameters of the network by verifying the change condition of a loss function and manually observing the effect of a generated image, and proving that the lower the value of the loss function is, the smaller the difference between the generated data and the real data is until time-varying data with better quality is generated.

The experiment adopts Mutual Information (MI), structural Similarity (SSIM), root Mean Square Error (RMSE) and quality evaluation to the B-scan data formed by the time-varying data. Wherein RMSE measures the error between the generated image and the original data, MI shows the amount of original data information contained in the generated data, SSIM reflects the structural information similarity, which defines B-scan as data with independent structural information, consisting of three factors of brightness, contrast, and structure. The mean value is used as an estimate of brightness, the standard deviation is used as an estimate of contrast, and the covariance is used as a measure of the degree of structural similarity. The evaluation indexes are defined as follows:

wherein, (x) _i ″,y _i "means (x) _i ′,y _i ') and n represents the total number of elements in the input data.

MI(X ₁ ,X ₂ )＝H(X ₁ )+H(X ₂ )-H(X ₁ ,X ₂ )

Wherein H (S) ₁ ,S ₂ ) And the joint information entropy of the input sequence B-scan is represented.

Wherein, the liquid crystal display device comprises a liquid crystal display device,is X ₁ Average value of>Is X ₂ Average value of>Is X ₁ Variance of->Is X ₂ Variance of->Is X ₁ And X ₂ Covariance of c ₁ ＝(k ₁ L) ² ，c ₂ ＝(k ₂ L) ² Is a constant used for stabilization, where L is the dynamic range of the element value and k ₁ ＝0.01，k ₂ =0.03. The structural similarity ranges from 0 to 1, and when the two data are identical, the value of SSIM is equal to 1.

(3) GPR time-varying data generation and quantization analysis

In order to verify the effectiveness of the model of this embodiment, this embodiment uses bicubic spline interpolation, HANTS interpolation, GAN network interpolation, VAE interpolation to perform data interpolation on the real GPR pipeline time-varying data set, normalizes the generated result to a pixel space of 0-255, and extracts local arrays of the same region a, b of the generated result of each method to perform amplification comparison, and the result is shown in fig. 9. The original data area a has a hyperbolic wave structure which is partially overlapped with the high-frequency clutter. As can be seen from fig. 9 (a), the high frequency clutter is still closely aligned in bicubic interpolation, and the structural information of the top part of the hyperbolic wave is blocked. Bottom clutter is not removed completely in the HANTS interpolation, and part of hyperbolic waves are shielded. The hyperbolic wave structure is obviously weakened in the VAE interpolation, and the structural information is not well preserved. In the GAN interpolation and the method of the embodiment, the hyperbolic wave structure is clearly characterized, and the characteristic information is well reserved. But the method of the embodiment has better inhibition effect on the high-frequency clutter, and most of the high-frequency clutter is removed.

There is a distinct hyperbolic wave structure in the original data region b and there is an irregular offset. As can be seen from fig. 9 (b), in bicubic interpolation, the problem of irregular offset in the hyperbolic wave structure is more serious, and the hyperbolic wave cannot be clearly distinguished. In both HANTS interpolation and VAE interpolation, the problem of irregular offset between data is alleviated, but the hyperbolic wave structure is significantly weakened. The GAN interpolation and the method of the embodiment better keep the structural characteristics of hyperbolic waves, the lamellar boundaries of different depths are obvious, and the irregular offset phenomenon is weakened. However, as shown in fig. 10, GAN interpolation is added with a significant irregular offset, indicating that the model is not stable.

The experimental results show that the hidden space data of the sequence time-varying array can be obtained by using the method of the embodiment, the hyperbolic wave characteristics can be clearly represented, the irregular offset phenomenon among the data is restrained, and the space data density is improved. Thereby verifying the validity and robustness of the present embodiment.

And generating a dense time-varying array by using different interpolation methods for the two types of data sets, extracting a real slice at the same position of the slice and the original sparse array from the dense time-varying array, performing quantization analysis, and respectively calculating average values of each result MI, each result RMSE and each result SSIM in the normalized pixel space, wherein the results are shown in the table 1. Higher MI values indicate greater continuity of information contained in the generated array, lower RMSE values indicate less error between the generated array and the original array, and higher SSIM values indicate more similarity in structural relationship between the generated array and the original array. From table 1, the time-varying array generated by the method of the present embodiment is superior to the other four methods in terms of mutual information, mean square error, structural similarity, etc. Compared with bicubic interpolation, the RMSE obtained by the method of the embodiment is reduced by 60% and the SSIM is improved by 48% in terms of numerical value. Compared with the HANTS interpolation method, the RMSE obtained by the method is reduced by 54%, and the SSIM is improved by 28%. It is described that the constraint embedded data generation model is effective in reducing irregular deviations. Compared with GAN interpolation, SSIM obtained by the method is improved by 4%, and RMSE is reduced by 29%. Compared with the VAE interpolation, the SSIM obtained by the method of the embodiment is improved by 14%, and the RMSE is reduced by 38%. The importance of the controller to dynamically adjust the loss weights for quality generation and learning ability is illustrated. The three indexes show that the quality of the reconstructed data generated by the method of the embodiment is higher, and the method is superior to other methods in the aspects of information correlation and structural continuity.

Table 1 comparison of objective evaluation indicators of different algorithms on reconstruction of different data sets

(4) Three-dimensional reconstruction experiment

(4-1) constructing a dense time-varying array of the real data set by adopting the data generation method based on the control vae in the embodiment, and performing data interpolation in a sparse array space. The generated data is controlled to tilt between different sampling sources by interpolating eigenvectors between two points in hidden space. As shown in fig. 11, in order to more intuitively embody the effectiveness of the method of the embodiment in improving the data density, the reconstruction process expands 1024 tracks of data contained in every two sequences of B-scan data in the sparse data set into 5120 tracks of data, that is, eight tracks of B-scan sections composed of pseudo-track data are interpolated in every two tracks of B-scan data, so as to obtain a C-scan-like three-dimensional dense time-varying array.

The space data density of the time-varying array after reconstruction is improved. And selecting fragments from the original sparse array and the reconstructed dense time-varying array, and performing three-dimensional visual analysis by an optical flow imaging processing algorithm, wherein the three-dimensional reconstruction result and the section display are shown in fig. 12 and 13. As can be seen from comparison between fig. 12 (b) and fig. 13 (b), the three-dimensional model constructed by the original sparse time-varying array has poor feature continuity, and the hyperbolic wave structure features are missing along with the running of the measuring line. The three-dimensional structure model obtained by reconstructing the dense time-varying array field through three-dimensional visualization by the method of the embodiment has clear background, the hyperbolic wave structure characteristics and obvious horizon structure information are reserved, and the structural continuity of the hyperbolic wave in the three-dimensional model is improved.

(4-2) constraint embedding effectiveness experiments

To verify whether the constraint embedded in the control vae model helps to eliminate irregular offset between data, the dense array generated by the control vae and the dense time-varying array generated by the method of the embodiment are directly used for three-dimensional visual comparison without constraint embedding processing, and the comparison result is shown in fig. 14. The hyperbolic wave crest is calibrated as an alignment point for visual visualization, and the result is shown in fig. 15. As can be seen from fig. 14 (a), there is a large amount of structural overlap in the hyperbolic wave portion without the embedded constraint, and errors gradually accumulate in the reconstruction process, so that the three-dimensional model cannot accurately reflect the actual distribution and structural trend of the underground target object. And fig. 14 (b) shows that the constrained embedded reconstructed dense time-varying array three-dimensional model has clear horizon background and high hyperbolic wave structure continuity.

As can be seen from fig. 15 (b), before constraint embedding, irregular offset between data causes unrealistic peak positions and even missing phenomena. The constraint embedding algorithm adopted in the embodiment can accurately detect and register the double Qu Bote sign areas, irregular offset among data is effectively eliminated by embedding constraint information, accumulated errors generated by reconstruction are reduced, and the reconstructed three-dimensional model has good structural continuity and better represents the distribution and trend of a target object.

The experimental result shows that in the construction of the dense time-varying array, the influence of irregular offset among data on data reconstruction can be effectively reduced through constraint embedding obtained through data registration, the space data density can be improved through a time-varying data generation network based on control VAE, the constructed dense time-varying array is small in error and high in precision, and the real distribution condition of an underground structure is effectively represented.

Example two

The embodiment discloses a dense time-varying array construction system based on a controllable variable self-encoder, comprising:

It should be noted that, the first acquiring module, the second acquiring module, and the dense time-varying array acquiring module correspond to the steps in the first embodiment, and the modules are the same as the examples and application scenarios implemented by the corresponding steps, but are not limited to the disclosure in the first embodiment. It should be noted that the modules described above may be implemented as part of a system in a computer system, such as a set of computer-executable instructions.

Example III

The third embodiment of the invention provides an electronic device, which comprises a memory, a processor and computer instructions stored on the memory and running on the processor, wherein the computer instructions complete the steps of the method for constructing the dense time-varying array based on the controllable variable self-encoder when the computer instructions are run by the processor.

Example IV

A fourth embodiment of the present invention provides a computer readable storage medium storing computer instructions that, when executed by a processor, perform the steps of the above-described dense time-varying array building method based on a controllable variable self-encoder.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The foregoing embodiments are directed to various embodiments, and details of one embodiment may be found in the related description of another embodiment.

The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. The method for constructing the dense time-varying array based on the controllable variable self-encoder is characterized by comprising the following steps:

2. The method for constructing a dense time-varying array based on a controllable variable self-encoder according to claim 1, wherein the performing time-varying data registration on the trace data in the two-dimensional profile data based on the region-associated feature constraint, and acquiring the additional feature information comprises:

3. The method for constructing a dense time-varying array based on a controllable variable self-encoder according to claim 1, wherein said inputting additional feature information and track data into a predetermined controllable variable self-encoder comprises:

4. A method of constructing a dense time-varying array based on a controllable variable self-encoder according to claim 3, wherein the joint distributed sampling of the low-dimensional vector and the constraint vector is specifically: inputting the low-dimensional vector and the constraint condition vector into an encoder, performing multi-element Gaussian distribution processing on the low-dimensional vector and the constraint condition vector, decomposing the input vector into a mean vector and a standard deviation vector which accord with multi-element Gaussian distribution, multiplying the standard deviation by random noise which accord with standard normal distribution by using a re-parameterization skill, and adding the mean to obtain normal distribution of reconstruction parameters; sampling from the reconstruction parameters and splicing with the condition vector to obtain the hidden vector containing both shallow information and deep features.

5. A method of constructing a dense time-varying array based on a controllable variable self-encoder as claimed in claim 3, wherein said interpolating samples between hidden vectors in two hidden spaces to form a new time-varying sequence synthesis vector is specifically: establishing a mapping between points of a priori distribution area in the hidden space and the input, wherein the mapping is realized through reconstruction loss of a decoder and hidden variable weight parameter constraint; and then interpolating and sampling between hidden vectors in the two hidden spaces to form a time-varying sequence synthesized vector.

6. The method for constructing a dense time-varying array based on a controllable variable self-encoder according to claim 1, wherein KL-divergence of each training step is sampled using a feedback control concept during training of the controllable variable self-encoder, and the super-parameter β (t) is adjusted accordingly by an iterative algorithm to stabilize the KL-divergence at a desired value.

7. The method of dense time-varying array construction based on a controllable variable self-encoder of claim 1, wherein the controllable variable self-encoder comprises a convolutional layer, a pooling layer, an encoder, and a decoder connected in sequence.

8. A dense time-varying array building system based on a controllable variable self-encoder, comprising:

9. An electronic device comprising a memory and a processor and computer instructions stored on the memory and running on the processor, which when executed by the processor, perform the steps of any of claims 1-7.

10. A computer readable storage medium storing computer instructions which, when executed by a processor, perform the steps of any of claims 1-7.